Performance of ML modulation classifier for multilevel QAM signals

In this paper, performance of the ML modulation classifier (MLMC) for multilevel QAM signals with respect to the number of observation samples is studied. A part of complex symbols are assumed to be shared by several modulation formats in common so that the smaller formats are subsets of the larger ones. As a result, correct classification ratio (CCR) of the MLMC algorithm is over 97 % in all SNR regions when the true constellation is 4 QAM. Even perfect classification can be achieved with triple-sample observation. In the case of the 16, 64 and 256 QAM, double-sample observation shows around 15 % increased CCR when compared to the single-sample observation scheme. And additional 10 % improvement can be obtained when three consecutive samples are exploited for the likelihood ratio tests. However, when the size of true constellation is increased, there are performance limits in single-and double-sample classification schemes. Such limits are mainly caused by the symbols shared by multiple constellations in common. Nevertheless, the MLMC with triple-sample observation shows almost perfect classification performance in high SNR region.